KerasCV is a repo of modular CV components that allow computer vision researchers to assemble state of the art computer vision pipelines --- but there are still open questions:
- what pretrained backbone is optimal for OD on real world datasets
- what augmentation pipeline is optimal for training most OD models
- what scores for MaP and Recall can be achieved using KerasCV
- are anchor free models (YOLOx) as powerful as anchor based models
This repo attempts to answer all of these questions through extensive empirical experimentation.
Current tasks implemented:
- PascalVOC
- Arthropods (results training now)